US issues Framework for Artificial Intelligence Diffusion

Background

The Framework for Artificial Intelligence Diffusion was an interim final rule issued by the US Department of Commerce in the final week of the Biden administration.^[2] Because it was issued in that form, it took effect on 13 January 2025.^[1] The rule set a separate compliance date of 15 May 2025, with some provisions delayed to 15 January 2026,^[1] and was reported as having originally been due to take effect on 15 May 2025.^[2]

Its subject was not a single product but a channel: the movement of advanced computing hardware and of the trained parameters of the largest AI models across borders.

What the rule required

The framework required licences for the export, re-export or in-country transfer of advanced computing integrated circuits and of the model weights of the most advanced AI models.^[1] Model weights are the numbers produced by training that determine how a model behaves; the rule added a new export control specifically covering them.^[1]

The rule also revised licence requirements and review policy for advanced integrated circuits, expanded the available licence exceptions, and added red-flag guidance on model weights — pointers meant to help exporters recognise transactions that deserve scrutiny.^[1] It further updated the Data Center Validated End User authorisation, a channel designed to ease exports of advanced computing items to destinations considered low risk.^[1]

Three tiers

Reporting by Securities Times described the framework as dividing the world into three tiers.^[2] Seventeen countries and Taiwan made up the first tier, which received unrestricted access to chips.^[2] About 120 countries fell into the second tier, where chip quantities were capped.^[2] The third tier — mainland China including Hong Kong and Macao, Iran, Russia and North Korea — was barred from importing advanced US AI chips.^[2]

Reaction

The rule drew immediate criticism globally.^[2] In mid-April, some US lawmakers wrote to the Commerce Secretary asking for it to be withdrawn.^[2]

Why it matters

Export controls have traditionally been written around physical goods: a machine, a component, a piece of equipment can be counted, inspected at a border and held back. A trained model is different. Its capability lives in a file of weights that can be copied and sent over a network, which is why the framework's extension of licensing to model weights, alongside the chips used to train and run models, is the part that made it unusual.^[1]

The tiered structure works the other way round from a simple ban. Instead of one list of restricted destinations, it sorts most of the world into categories with different levels of access: unrestricted for a small group of close partners, capped quantities for a large middle group, and a prohibition for a short list of countries.^[2] That design turns an export rule into a map of who may buy how much, and it attaches conditions to countries that are neither close allies nor outright targets.

The rule also shows how quickly the technical details of AI policy became administrative ones. Licence requirements, exceptions, red-flag guidance and validated end-user authorisations are the ordinary machinery of export administration; applying them to model weights and to data-centre customers pushed AI governance into that machinery rather than leaving it to statements of principle.^[1]

The framework's timing mattered as much as its content. Issued in the last week of one administration with a compliance date months later,^[1][2] it left room for the debate that followed, in which critics abroad and lawmakers at home pressed for withdrawal.^[2]